Daily crew schedules that took hours by hand now start from a recommendation.

Contexture is the largest commercial window installation company on the US East Coast, and it was growing quickly, in large part through acquisition. Each deal added installation crews in new regions working new job sites, and deciding where to send those crews each day became a logistics problem that grew with every acquisition.

A workable schedule had to satisfy a long list of rules at once. Jobs had hard start and end dates set by the general contractor’s readiness and the contract. Every crew needed a foreman, apprentices couldn’t work without one, and installers had to hold the skill level and safety certifications a job called for. Regions had their own work-day limits and union hours. Boston jobs carried resident and minority labor percentages that had to be met. The company had been building schedules by hand against all of this. Sites slipped, deliveries moved, and crews called out, so on a busy day the lead scheduler rebuilt the schedule as many as eight times, hours of work in all. As crews and sites multiplied, that had become one of the limits on how fast the business could expand.

We worked with the Contexture team to gather every rule a schedule had to respect and sorted them into three kinds: hard constraints the system can never break, soft ones it follows when it can, and optimization goals it pursues once the first two are met. That came to 24 rules in all. No off-the-shelf scheduling product could hold them, so we built a recommendation engine on Google OR-Tools, the open source optimization suite Google uses for its own scheduling and routing, and connected it to the installer data Contexture already kept: skills by product code, foreman eligibility, region, availability, certifications, and current assignments.

The engine evaluates all of the constraints simultaneously and returns a full day’s schedule. Its first optimization priority is continuity, keeping crews on sites they already know, followed by deadline pressure, job complexity, skill match, travel distance, and daily utilization. It is a recommender, so the scheduler accepts, declines, or adjusts what it proposes. A job the scheduler has already assigned by hand stays put and the engine fills in around it. When there aren’t enough installers to cover the jobs that must run, or no available installer meets a job’s minimum skill, it says so instead of quietly producing a bad answer. To protect the budget we structured the build in three phases, beginning with a minimum viable product, and the CEO chose to fund all three at once.

For the final phase we ran the recommender nearly every day for more than two months against the schedules Contexture’s lead scheduler produced independently, and by handoff it agreed with her decisions 90 to 95% of the time. The work that had taken her hours a day, rebuilt as many as eight times as conditions changed, is now produced by the recommender. Her part is the adjustments it can’t make from data, such as specialized equipment, material deliveries, travel, and who works well with whom, which in the Northeast came to one or two overrides a week.

The larger gain is what this does for the acquisition strategy. Each new region had meant more scheduling work and, before long, a scheduling team to carry it. The recommender takes on a new region’s crews and rules as data, so Contexture can keep buying and integrating installers without building that team. Since launch, the lead scheduler has had enough time back to take on broader needs of the business.